@ferdy You could equally claim that 'Go has already existed before it, that is an evidence that Hand and Brain is a derivation from Go'. Suggesting that Dice Chess is so similar to Hand & Brain that we can confidently say the latter idea came from the former is laughable. In any case, the idea of 'one player picks pieces, the other one moves them' (a die is not a player: it is a move-by-nature) is what we want to know the origin of, as well as when it was first called 'Hand & Brain' etc. See Pritchard for what historical description of chess variant origins looks like

By Tandem Chess I refer not to Bughouse but to Tandem/Alternation chess, which again I mentioned in my first post, where the players take turns moving (on the same board). Pritchard identified the origins of this variant. This is not mentioned on Wikipedia and I don't know that it is more popular than Hand & Brain.

@DialFrost Presumably you'd say the same about Tandem Chess, but Pritchard was able to identify its history back to an 1875 journal. There's no reason to believe it's impossible to do similar for Hand & Brain

I'm not asking for a listing of all the mentions you can find online of Hand & Brain. I'm asking about the origins of the variant. It's obvious from Kosteniuk's blog that the variant was established by 2013 when she refers to it, likely well before, for the reasons I already explained in my post (clearly both the name and exact practice, which she refers to as a casual consensus, were well-established by the time of writing). For an example of what proper historical descriptions of the origins of variants look like, have a read through Pritchard's Classified Encyclopedia (free online).

@ferdy There is no evidence that the (very abstract and somewhat tenuous) connection you draw between the ideas in Dice Chess and Hand & Brain inspired or gave rise to Hand & Brain; indeed it seems unlikely. (But in any case, it is substantive evidence we are looking for.)

@SteveBennett I imagine note-taking is not admissible. That being the case, human memory capacity is very limited and the answer would be quickly reached (let's say, a couple of hours per move). If note-taking is possible, the next limit would be motivation to keep exploring different lines for months per move, as you say. However, working 24/7, I doubt humans would be motivated to perform well analysing months per move trying to collate all notes on paper and make sense of them after. A few days per move is probably still the max I'd guess.

@BrianDrake Castling is the exception here: the heuristic is excellent in general. The castling-specific rules are oddities somewhat, nor does the fact that you can't castle out of check (or through check) help understand why one must move out of check when they could deliver mate instead.

The question isn't asking about humans (or even computers, really, though the strongest computers are bound to be the best proxy), it's asking what the theoretical result is. In other words, when you say stuff like "most likely white cannot win consistently" or "there is a chance this can go down to 0.0", you are asked to give a percentage.

@Allure Firstly, which plot are you looking at in the link to deduce that 25% of games at +0.9 eval are won? Secondly, it's worth noting that we have to understand what percentage are won by Stockfish when the true eval (or, let's say, high-depth eval after long time thinking) is +0.9 or +1.2, not when the eval Stockfish sees while playing bullet is +0.9 or +1.2

@ferdy It's clear that this doesn't go far towards answering the original question (whether or not the position is a theoretical win or draw)

@HaukeReddmann I'm not sure I get it. My point was that, if you do want whitebox heuristics, they are likely to require considerations of computer evaluations and/or the game-theoretic nature of the moves chosen (how exactly Kasparov tries to push for a complicated/difficult position, as opposed to Carlsen). If you are happy with blackbox methods for individuals' playstyles then this has already been done: cs.toronto.edu/~ashton/pubs/maia-personalized2021.pdf

Why it is relevant: well, I suspect that maybe easy statistical heuristics like your example wouldn't cover it, and you'll need to rely on some game-theoretic measures regarding evaluations, how exactly they try to make the position harder for their opponent (to find evaluation-preserving lines in), etc.

what do you mean "such differences would be moot for a hypothetical 32-piece tablebase"? I think that if you had such a tablebase, you could not only always win (against humans) but also choose what playstyle you want to win with.

cf chess.stackexchange.com/questions/8856/… Kaufman (creator of Komodo and founding father of engine chess) did some experiments to test how much of a disadvantage not being able to castle is for computers. TLDR it's a pretty big one

@SecretAgentMan The fact that 'self-learning' (i.e. learnt by self-play) engines decide to castle is precisely what IanRingrose is citing as evidence that castling is often advantageous

Another piece of evidence to consider maybe: material odds matches (the only kind of competitive battle humans still have with engines). Kaufman has written a lot on what results Komodo should expect against humans of given rating at given time-controls. As a quick summary, it's much easier for Komodo to recover a substantial material deficit (like a Knight) at fast time-controls. (FWIW, currently Komodo is just short of par at rapid controls with an average 2500 GM after giving Knight odds.)

TLDR (thanks to oscarsmith for pointing it out) seems to be that Leela was an infant then and not a 'real engine'. At a 3250 rating on Lichess it should be substantially weaker than even Rybka (which might be 3250+ in FIDE classical on a human scale, and almost certainly better at blitz than slow controls), let alone any contemporary engine. It received opportunities for play solely on the basis of being an imitation effort of AlphaZero. We can guess Leela has absorbed thousands of rating points since the match with Andrew Tang.

I had an inkling the reason was something like that. (IIRC a slightly later prototype version of Leela scored only 0-2 draws in the bottom division of TCEC. This was just an early experiment in imitating AlphaZero then, not the serious competitor engine it is now. ) Then that forms an important part of an answer I should think